Personalized learning resource recommendation method based on subject knowledge graph and graph neural network

Abstract

The invention relates to a personalized learning resource recommendation method based on a subject knowledge graph and a graph neural network, and belongs to the field of education content recommendation. The method comprises the following steps: extracting key entities in a textbook and a relationship between the key entities, and constructing a subject knowledge graph; establishing a preliminary learning record table according to the behavior data of the student in the learning process; through knowledge tracking model integration, predicting mastery degree distribution of students for each knowledge point in the learning record sheet, and constructing student portraits; embedding the knowledge points in the subject knowledge graph and the association relationship thereof into the student portrait, and updating the knowledge embedding representation in the student portrait through a graph neural network; and constructing a personalized recommendation model, and pushing learning resources conforming to the current learning state of the student by adopting a mixed recommendation method of collaborative filtering and content recommendation. According to the method, personalized demands of students can be accurately captured in multiple dimensions, and the accuracy, generalization and interpretability of recommended contents are effectively improved.

Description

Technical Field

[0001] The present invention belongs to the field of educational content recommendation and relates to a personalized learning resource recommendation method based on subject knowledge graph and graph neural network. Background Art

[0002] Personalized recommendation methods based on subject knowledge graphs have been widely used in fields such as educational content recommendation, and are particularly suitable for scenarios such as personalized learning systems and intelligent education platforms. Traditional recommendation methods usually make recommendations based on students' historical behavior data, lack a comprehensive understanding of students' knowledge structure, and have difficulty capturing students' deep-level learning needs and knowledge mastery, resulting in insufficient accuracy and pertinence of recommended content. In addition, existing methods have limited understanding of the correlation between knowledge points, and the recommendation results lack contextual information support, which makes it difficult for the system to provide sufficient generalization and interpretability in a multi-knowledge point environment.

[0003] To solve the above problems, knowledge enhancement strategies based on knowledge graphs have gradually become an effective means to improve the performance of personalized recommendation systems. Knowledge enhancement strategies can not only introduce rich semantic information between knowledge points in student portraits, but also propagate knowledge graphs at multiple levels through graph neural networks, thereby enhancing the system's ability to understand students' knowledge mastery and improving the contextual perception level of the recommendation system and the diversity and accuracy of recommended content. Existing personalized recommendation methods based on knowledge graphs have two main limitations: on the one hand, some methods have insufficient information fusion in combining knowledge graph embedding and student portraits; on the other hand, in terms of dynamically updating student portraits and recommendation results, the real-time performance and flexibility of the system still need to be improved. Summary of the invention

[0004] In view of this, the purpose of the present invention is to provide a personalized learning resource recommendation method based on subject knowledge graph and graph neural network. The personalized recommendation method of subject knowledge graph based on knowledge enhancement combines multi-knowledge point fusion, knowledge enhancement strategy and student portrait, which improves the accuracy, generalization and interpretability of the system, and is suitable for deep personalized recommendation in the field of intelligent education.

[0005] In order to achieve the above object, the present invention provides the following technical solutions:

[0006] A personalized learning resource recommendation method based on subject knowledge graph and graph neural network, the method comprising the following steps:

[0007] S1. Extract key entities and the relationships between key entities in the textbooks and construct a subject knowledge graph;

[0008] S2. Establish a preliminary learning record sheet based on students’ behavioral data during the learning process;

[0009] S3. Integrate and predict the distribution of students’ mastery of each knowledge point in the learning record table through the knowledge tracking model to build a student portrait;

[0010] S4, embed the knowledge points and their associations in the subject knowledge graph into the student portrait, and update the knowledge embedding representation in the student portrait through the graph neural network;

[0011] S5. Build a personalized recommendation model and use a hybrid recommendation method of collaborative filtering recommendation based on group characteristics and content recommendation based on semantic features to push learning resources that are in line with students' current learning status.

[0012] Furthermore, in step S1, key entities and their relationships in the textbook are extracted by manual extraction and semi-automatic extraction, and the entities and relationships are structured and stored in the form of entity-relationship-entity or attribute to construct a subject knowledge graph. The subject knowledge graph contains a network structure of various knowledge points and their associations, and the subject knowledge graph is stored in the Neo4j graph database.

[0013] Further, in step S2, the behavior data of the students in the learning process is collected, and the behavior data includes the exercises completed by the students, the knowledge points related to the exercises, the difficulty of the exercises, and the answer results;

[0014] Preprocess and normalize the students' behavior data to generate a preliminary learning record table, where preprocessing includes standardization and clearing operations to remove duplicate record data and missing values;

[0015] The learning records in the preliminary learning record table are classified by characteristics, and the learning records are divided into different types of data according to the frequency of answering questions, the difficulty of answering questions, and the degree of mastery of knowledge points.

[0016] Further, in step S3, the knowledge tracking model includes an embedding layer, a feature concatenation layer, a fully connected layer, a deep neural network layer and an output layer, which includes the following steps:

[0017] S31. Convert the knowledge point encoding into vector representation through the embedding layer:

[0018] e s =Eebedding s (s)

[0019] e q =Eebedding q (q)

[0020] e k =Eebedding k (k)

[0021] ea =Eebedding a (a)

[0022] In the formula, e s 、e q 、e k and e a are the embedding vectors of students, questions, knowledge points and answers respectively. s, q, k, a are the data information corresponding to students, questions, knowledge points and answers respectively. s (), Eebedding q (), Eebedding k (), Eebedding a () are the embedding mapping functions corresponding to students, questions, knowledge points and answers respectively;

[0023] S32, concatenate the embedded vectors into an overall feature vector x through the feature concatenation layer:

[0024] x=[e s , e q , e k , e a ]

[0025] S33, pass x as input to the fully connected layer, integrate the knowledge point embedding and student features through the fully connected layer, fuse the features of multiple knowledge points into one, and generate a feature vector suitable for student knowledge status analysis; input the fused feature vector into the deep neural network to predict the student's mastery probability on each knowledge point:

[0026] KLinear(x)=W base ·σ(x)+W spline ·B(x)

[0027]

[0028] Where KLinear(x) is the linear calculation result of the fully connected layer on the input x, that is, the output value of the fully connected layer, and W base is the basis weight, σ(x) is the activation function of the basis weight; W spline is the spline weight, B(x) is the basis function of the spline interpolation; k and s are the current node index and the order of the spline interpolation, respectively, grid k Represents the value of the kth grid node used for spline interpolation in the fully connected layer, grid k+s+1 Represents the k+s+1th grid node, which is used to calculate the distance between the current node and the next node. k+s+2 Indicates the k+s+2th grid node, which is used for higher index values ​​required in interpolation calculations. k+s-2Represents the k+s-2th grid node, which is used to calculate earlier node values ​​and to construct the polynomial basis during interpolation. k+1 Represents the k+1th grid node, which is used in the calculation of the interpolation basis function to determine the boundary of the interpolation range;

[0029] S34, the output layer uses the Sigmoid activation function to generate the predicted probability y of students mastering the knowledge points:

[0030] y=σ(KLinear3(KLinear l (KLinear1(x))))

[0031] Based on the students' knowledge mastery distribution, we can obtain a student portrait that reflects their learning status, including the mastery level and progress of each knowledge point.

[0032] Further, in step S4, the following steps are included:

[0033] S41. Use graph embedding method to vectorize each knowledge point entity in the knowledge graph: vectorize the knowledge point k and the relationship between the knowledge point and other knowledge points r ij This is processed through a language model, which encodes the input text description into an embedded representation:

[0034] e k =BERT(T k )

[0035]

[0036] Where, T k is the text description of knowledge point k; It is a relationship ij Text description, BERT() is a pre-trained bidirectional encoder representation converter model used to convert input text into an embedding vector;

[0037] S42. Integrate the knowledge graph embedding result into the student portrait, including: based on the relational structure in the knowledge graph, use the graph neural network to propagate the semantic information of the knowledge point, and integrate the upstream and downstream information of the relevant knowledge point into the current student portrait; perform feature splicing on the student's mastery of the knowledge point and the vector embedded in the knowledge graph to form a more complete representation of the student's knowledge state, wherein, combined with the neighborhood aggregation operation of the graph neural network, the embedding representation of the knowledge point and the relationship is integrated into the student portrait, and the student portrait embedding after knowledge enhancement is obtained:

[0038] For each knowledge point node k, collect its neighboring nodes The embedded information of , the neighborhood information is integrated through the aggregation function f:

[0039]

[0040] Where W(l) is the weight matrix of the lth layer, σ() is the activation function, concat() is the concatenation function, and f() is the aggregation function, which aggregates the embedded information of node k and its neighboring nodes. is the embedding information of the t-th domain node of node k in the l-th layer, and Respectively represent the embedding information of node k in layer l and layer l+1;

[0041] The knowledge graph information after graph embedding is used to supplement the background knowledge of the student portrait and identify the mastery relationship of students on knowledge points.

[0042] Furthermore, in step S5, the personalized recommendation model adopts a hybrid recommendation method to recommend learning resources by combining collaborative filtering recommendation based on group characteristics and content recommendation based on semantic characteristics. ik The calculation method is:

[0043]

[0044] In the formula, is the recommendation degree of knowledge point k given by collaborative filtering to student i is the recommendation degree of knowledge point k given by the content recommendation method to student i, and α is the weight parameter used to balance the influence of the two recommendation methods;

[0045] Based on the output of the personalized recommendation model, personalized learning content is pushed to students. The learning content includes: knowledge point materials, specific exercises and learning paths.

[0046] Furthermore, in the personalized recommendation model, the recommendation process of the collaborative filtering recommendation method based on group characteristics is as follows:

[0047] Based on the students’ knowledge and interest characteristics, we recommend other knowledge points that are similar to their known knowledge points. We calculate the similarity by vectorizing the knowledge points and students’ learning records:

[0048]

[0049] In the formula, i, j represent different students respectively; R ik , R jk Respectively represent the scores of student i and student j on knowledge point k, and n is the total number of knowledge points;

[0050] Predict the recommendation degree of student i for unlearned knowledge point k based on similarity

[0051]

[0052] In the formula, r i is the average mastery of student i, r jk is the mastery of knowledge point k by student i, and N(i) is the set of students most similar to student i.

[0053] Furthermore, in the personalized recommendation model, the recommendation process of the content recommendation method based on semantic features is as follows:

[0054] Use vector representation and semantic matching to achieve semantic enhancement of knowledge graph. Suppose knowledge point k is embedded in the graph embedding model of knowledge graph to obtain embedding vector v k , and the embedding vector of the student portrait is v i , the semantic similarity is calculated as follows:

[0055]

[0056] Where D is the vector dimension, v id Represents the student portrait embedding vector v i The d-th dimension vector in, v kd Represents the knowledge graph embedding vector v k The d-th dimension vector in ;

[0057] Based on the semantic similarity, the recommendation degree of student i for knowledge point k is predicted

[0058]

[0059] In the formula, β is a regulation parameter used to adjust the impact of content recommendation on the overall recommendation degree.

[0060] The beneficial effects of the present invention are:

[0061] The present invention generates a student's knowledge mastery portrait by constructing a multi-knowledge point fusion model, and uses the rich semantic information of the knowledge graph to enhance knowledge, integrating knowledge points and their associations into the student portrait. The semantic propagation mechanism based on the graph neural network further enhances the system's ability to understand the context of the knowledge graph. By combining student portraits with knowledge enhancement information, the present invention can accurately capture students' personalized needs in multiple dimensions, effectively improve the accuracy, generalization and interpretability of recommended content, and is suitable for deep personalized recommendations in scenarios such as education and content recommendation.

[0062] Other advantages, objectives and features of the present invention will be described in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below in conjunction with the accompanying drawings, wherein:

[0064] Figure 1 A flowchart of a personalized learning resource recommendation method based on subject knowledge graph and graph neural network of the present invention;

[0065] Figure 2 Schematic diagram of the recommendation process of the personalized learning resource recommendation method based on subject knowledge graph and graph neural network of the present invention. DETAILED DESCRIPTION

[0066] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0067] Among them, the drawings are only used for illustrative explanations, and they only represent schematic diagrams rather than actual pictures, and should not be understood as limitations on the present invention. In order to better illustrate the embodiments of the present invention, some parts of the drawings may be omitted, enlarged or reduced, and do not represent the size of actual products. For those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings may be omitted.

[0068] The same or similar numbers in the drawings of the embodiments of the present invention correspond to the same or similar parts; in the description of the present invention, it should be understood that if the terms "upper", "lower", "left", "right", "front", "rear", etc. indicate the orientation or position relationship, they are based on the orientation or position relationship shown in the drawings, which is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operate in a specific orientation. Therefore, the terms describing the position relationship in the drawings are only used for illustrative purposes and cannot be understood as limiting the present invention. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to specific circumstances.

[0069] See also Figure 1-2 , which is a personalized learning resource recommendation method based on subject knowledge graph and graph neural network.

[0070] Example

[0071] This embodiment provides a detailed process of a personalized learning resource recommendation method based on subject knowledge graph and graph neural network, such as Figure 1 As shown in the figure, it first extracts the key entities and the relationships between key entities in the textbook to build a subject knowledge graph; then collects and analyzes student behavior data, builds student portraits through the knowledge tracking model, and uses knowledge enhancement strategies to propagate knowledge graph information; finally, it builds a personalized recommendation model and provides a personalized recommendation strategy, which specifically includes the following steps:

[0072] Step 1: Extract the key entities and the relationships between key entities in the textbook, obtain the data form of entity-relationship-entity (or attribute), and construct the knowledge graph of the subject area;

[0073] The key entities and their relationships in the textbooks are extracted through manual extraction and semi-automatic extraction technology, and these entities and relationships are structured and stored in the form of entity-relationship-entity (or attribute) to construct a knowledge graph in the subject area. This knowledge graph contains the network structure of each knowledge point and its relationship. The knowledge graph in the subject area is constructed by processing the entities and relationships extracted from the subject textbooks, and stored in the Neo4j graph database. The entities and relationships in the subject knowledge graph are used to make personalized recommendations to students based on semantic logic and association rules.

[0074] Step 2: Collect students’ behavioral data during the learning process, including completed exercises, relevant knowledge points, exercise difficulty, answer results, and other information, pre-process and normalize the data, generate a preliminary learning record sheet for students, and classify and organize the data according to students’ behavioral characteristics to provide basic data for the subsequent construction of student portraits;

[0075] In the process of students using the learning system, the learning behavior data of students are collected. For example, the exercises that students have done, and the detailed information of various exercises that students have completed in the learning process are recorded; each exercise will have related knowledge point information, so that each exercise matches the specific knowledge point; collect students' performance when answering questions, and record the student's answer behavior corresponding to each exercise, including whether the answer is correct or not, with a correct answer recorded as 1 and an incorrect answer recorded as 0; record the difficulty level of the exercises as auxiliary data for evaluating students' learning ability.

[0076] Standardize and clean the collected behavioral data to ensure data integrity and consistency; remove duplicate records and missing values ​​to ensure that each record in the data set is unique and complete. Based on the preprocessed data, obtain a preliminary learning record table for each student, covering the student's behavioral performance on specific knowledge points, such as answering questions, learning progress, etc. Classify the student's learning records by behavioral characteristics and divide them into different types of data such as answering frequency, answering difficulty, and knowledge point mastery. This step will help to build student portraits later.

[0077] Step 3: Based on the student's learning record sheet, a multi-knowledge point fusion model is constructed to integrate and calculate the various knowledge points involved by the student, and a deep neural network is used to predict the student's mastery probability of each knowledge point, generating an accurate distribution of the student's knowledge mastery, thereby constructing a personalized portrait of the student, and comprehensively reflecting the student's learning status in different knowledge areas;

[0078] Then, based on the learning record table obtained in step 2, a knowledge tracking model is designed to integrate students’ data on multiple knowledge points. The knowledge point encoding is converted into a vector representation through the embedding layer to provide data with consistent format for subsequent deep learning network input:

[0079] e s =Eebedding s (s)

[0080] e q =Eebedding q (q)

[0081] e k =Eebedding k (k)

[0082] e a =Eebedding a (a)

[0083] In the formula, e s 、e q 、e k and e a are the embedding vectors of students, questions, knowledge points and answers respectively. s, q, k, a are the data information corresponding to students, questions, knowledge points and answers respectively. s (), Eebedding q (), Eebedding k (), Eebedding a () are the embedding mapping functions corresponding to students, questions, knowledge points and answers respectively.

[0084] Through feature concatenation, the embedded vectors are concatenated into an overall feature vector x:

[0085] x=[e s , e q , e k , e a ]

[0086] Pass x as input to the fully connected layer, integrate the knowledge point embedding and student features through the fully connected layer, fuse the features of multiple knowledge points into one, and generate a feature vector suitable for analyzing the student's knowledge status; the fused feature vector is input into the deep neural network to predict the student's mastery probability of each knowledge point.

[0087] KLinear(x)=W base ·σ(x)+W spline ·B(x)

[0088]

[0089] Where KLinear(x) is the linear calculation result of the fully connected layer on the input x, that is, the output value of the fully connected layer, and W base is the basis weight, σ(x) is the activation function of the basis weight; W spline is the spline weight, B(x) is the basis function of the spline interpolation; k and s are the current node index and the order of the spline interpolation, respectively, grid k Represents the value of the kth grid node used for spline interpolation in the fully connected layer, grid k+s+1 Represents the k+s+1th grid node, which is used to calculate the distance between the current node and the next node. k+s+2 Indicates the k+s+2th grid node, which is used for higher index values ​​required in interpolation calculations. k+s-2 Represents the k+s-2th grid node, which is used to calculate earlier node values ​​and to construct the polynomial basis during interpolation. k+1 Represents the k+1th grid node, which is used in the calculation of the interpolation basis function to determine the boundary of the interpolation range.

[0090] The output of the model passes through the Sigmoid activation function to generate the predicted probability y of the student's mastery of the knowledge point:

[0091] y=σ(KLinear3(KLinear2(KLinear1(x))))

[0092] Based on the distribution of students’ knowledge mastery, we can obtain a portrait of their learning status, including the mastery level and progress of each knowledge point. This portrait provides comprehensive and detailed student learning information support for subsequent knowledge enhancement and personalized recommendations.

[0093] Step 4: Introduce semantic information in the knowledge graph, embed knowledge points and their associations into student portraits, use graph embedding methods to convert entities and relationships in the knowledge graph into vector representations, combine with student portraits, enrich the background information in student portraits, and improve the model's understanding of student knowledge structure. In addition, apply graph neural networks on the knowledge graph to achieve multi-layer propagation of knowledge graph information, and further enhance the deep knowledge expression ability of student portraits by aggregating neighborhood information of knowledge points and dynamically updating knowledge embedding representations in student portraits;

[0094] Using the graph embedding method, each knowledge point entity in the knowledge graph is vectorized to retain its semantic features in the knowledge graph; each relationship in the knowledge graph is converted into a vector so that the relationship between knowledge points can be represented in the embedding space, helping the model understand the relationship between knowledge points.

[0095] Assume knowledge point k and its relationship with other knowledge points r ij It can be processed by a language model. The language model encodes the input text description into an embedded representation:

[0096] e k =BERT(T k )

[0097]

[0098] Where, T k is the text description of knowledge point k; It is a relationship ij BERT() is a pre-trained bidirectional encoder representation converter model used to convert input text into an embedding vector.

[0099] Combine the knowledge graph embedding results into the student portrait to enhance the knowledge representation ability of the portrait. Based on the relational structure in the knowledge graph, use the graph neural network to propagate the semantic information of knowledge points, and integrate the upstream and downstream information of related knowledge points into the current student portrait to enhance the understanding of the knowledge mastery structure. Feature splicing of students' mastery of knowledge points and vectors embedded in the knowledge graph to form a more complete representation of students' knowledge status.

[0100] When embedding knowledge points and relationships into student portraits, we can further combine the neighborhood aggregation operation of the graph neural network to obtain the knowledge-enhanced student portrait embedding. For each knowledge point node k, collect its neighborhood nodes The embedded information of , these neighborhood information are integrated through the aggregation function f:

[0101]

[0102] Where W(l) is the weight matrix of the lth layer, σ() is the activation function, concat() is the concatenation function, and f() is the aggregation function, which aggregates the embedded information of node k and its neighboring nodes. is the embedding information of the t-th domain node of node k in the l-th layer, and They represent the embedding information of node k at layer l and layer l+1 respectively.

[0103] The knowledge graph information after graph embedding supplements the background knowledge of the student portrait, enabling the model to identify the mastery relationship of students on knowledge points. For example, through the predecessor relationship, it can be judged that the student's mastery of knowledge point A is helpful to master knowledge point B, which enhances the semantic background and related knowledge understanding of the portrait. The knowledge-enhanced student portrait provides a richer personalized knowledge structure for the subsequent recommendation model, improving the system's understanding and utilization of the student's knowledge structure in the recommendation process.

[0104] At this time, the student portrait not only reflects the current knowledge mastery, but also contains contextual information about the student's potential knowledge needs, which enhances the recommendation model's ability to understand the implicit relationship between students' interest points, making personalized recommendations more accurate and effective.

[0105] Step 5: Combine the student portrait containing knowledge enhancement information with student needs to build a personalized recommendation model. This model uses a hybrid recommendation method that combines collaborative filtering recommendations based on group characteristics and content recommendations based on semantic features. Through collaborative filtering, the system can analyze student groups with similar learning characteristics and mine knowledge points that students may be interested in; content-based recommendation methods use semantic information in knowledge graphs to help the system understand students' specific knowledge needs. The above recommendation methods are used to recommend personalized learning resources, including personalized knowledge points, exercises, and learning paths, and push learning resources that match students' current learning status, helping students develop a learning path that suits them, thereby more effectively improving learning outcomes.

[0106] By analyzing the learning behavior, knowledge mastery, knowledge graph relationships and other features recorded in the student portrait, this information is matched with the student's current learning goals and needs. Student needs may include the need to master specific knowledge points, difficulty preferences, etc. The multidimensional characteristics of students are introduced into the model, including their knowledge mastery, learning behavior characteristics, knowledge point preferences, learning paths, etc. These features help to fully understand students' needs. Based on the student portrait, the correlation between the knowledge points in the knowledge graph is used to screen out content that is highly correlated with the student's current knowledge status. For example, by analyzing the adjacency relationship of related knowledge points through the knowledge graph, knowledge points that students have not yet mastered but are important and relevant are included in the recommendation range.

[0107] This method recommends other knowledge points similar to the students' known knowledge points based on their knowledge mastery and interest characteristics. The recommendation is made by calculating the similarity (such as cosine similarity) based on the vectorized results of the knowledge points and the students' learning records:

[0108]

[0109] In the formula, i, j represent different students respectively; R ik , R jk They represent student i’s score on knowledge point k, student j’s score on knowledge point k, and n is the total number of knowledge points.

[0110] Based on the similarity, we can predict the recommendation degree of student i for the unlearned knowledge point k

[0111]

[0112] In the formula, r i is the average mastery of student i, r jk is the mastery of knowledge point k by student i, and N(i) is the set of students most similar to student i.

[0113] In the content recommendation part, vector representation and semantic matching are used to achieve semantic enhancement of knowledge graph. Assume that knowledge point k is embedded with vector v through the graph embedding model of knowledge graph. k , and the embedding vector of the student portrait is v i , the semantic similarity is calculated as follows:

[0114]

[0115] Where D is the vector dimension, v id Represents the student portrait embedding vector v i The d-th dimension vector in, v kd Represents the knowledge graph embedding vector v k The d-th dimension vector in ;

[0116] Based on the semantic similarity, the recommendation degree of student i for knowledge point k is predicted

[0117]

[0118] In the formula, β is a regulation parameter used to adjust the impact of content recommendation on the overall recommendation degree.

[0119] Combining the results of collaborative filtering and content recommendation, we can get the comprehensive recommendation degree r of the knowledge point. ik :

[0120]

[0121] In the formula, α is a weight parameter used to balance the influence of the two recommendation methods.

[0122] A personalized recommendation model is constructed by combining the results of student portraits and knowledge point association analysis. Through this model, the knowledge point features and multi-dimensional needs in the student portraits are input into the recommendation algorithm, and a set of learning content that best matches the student's current learning stage and needs is output. Finally, based on the output of the personalized recommendation model, personalized learning content is pushed to students, including knowledge point materials, specific exercises, and learning paths, etc., to ensure that the recommended content is in line with the student's knowledge mastery level and meets the student's learning interests and needs.

[0123] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution, which should be included in the scope of the claims of the present invention.

Claims

1. A personalized learning resource recommendation method based on subject knowledge graph and graph neural network, characterized by: The method comprises the following steps: S1. Extract key entities and the relationships between key entities in the textbooks and construct a subject knowledge graph; S2. Establish a preliminary learning record sheet based on students’ behavioral data during the learning process; S3. Integrate and predict the distribution of students’ mastery of each knowledge point in the learning record table through the knowledge tracking model to build a student portrait; S4, embed the knowledge points and their associations in the subject knowledge graph into the student portrait, and update the knowledge embedding representation in the student portrait through the graph neural network; S5. Build a personalized recommendation model and use a hybrid recommendation method of collaborative filtering recommendation based on group characteristics and content recommendation based on semantic features to push learning resources that are in line with students' current learning status.

2. According to claim 1, a personalized learning resource recommendation method based on subject knowledge graph and graph neural network is characterized by: In step S1, the key entities and their relationships in the textbook are extracted by manual extraction and semi-automatic extraction, and the entities and relationships are structured and stored in the form of entity-relationship-entity or attribute to construct a subject knowledge graph. The subject knowledge graph contains a network structure of each knowledge point and its relationship, and the Neo4j graph database is used to store the subject knowledge graph.

3. According to claim 2, a personalized learning resource recommendation method based on subject knowledge graph and graph neural network is characterized by: In step S2, the behavior data of students in the learning process is collected, and the behavior data includes exercises completed by students, knowledge points related to the exercises, difficulty of the exercises, and answer results; Preprocess and normalize the students' behavior data to generate a preliminary learning record table, where preprocessing includes standardization and clearing operations to remove duplicate record data and missing values; The learning records in the preliminary learning record table are classified by characteristics, and the learning records are divided into different types of data according to the frequency of answering questions, the difficulty of answering questions, and the degree of mastery of knowledge points.

4. According to claim 3, a personalized learning resource recommendation method based on subject knowledge graph and graph neural network is characterized by: In step S3, the knowledge tracking model includes an embedding layer, a feature concatenation layer, a fully connected layer, a deep neural network layer, and an output layer, which includes the following steps: S31. Convert the knowledge point encoding into vector representation through the embedding layer: and s =Eebedding s (s) e q =Eebedding q (q) and k =Eebedding k (k) e a =Eebedding a (a) In the formula, e s 、e q 、e k and e a are the embedding vectors of students, questions, knowledge points and answers respectively. s, q, k, a are the data information corresponding to students, questions, knowledge points and answers respectively. s ( ), Eebedding q ( ), Eebedding k ( ), Eebedding a ( ) are the embedding mapping functions corresponding to students, questions, knowledge points and answers respectively; S32, concatenate the embedded vectors into an overall feature vector x through the feature concatenation layer: x=[and s ,And q ,And k ,And a ] S33, pass x as input to the fully connected layer, integrate the knowledge point embedding and student features through the fully connected layer, fuse the features of multiple knowledge points into one, and generate a feature vector suitable for student knowledge status analysis; input the fused feature vector into the deep neural network to predict the student's mastery probability on each knowledge point: KLinear(x)=W base ·σ(x)+W spline ·B(x) Where KLinear(x) is the linear calculation result of the fully connected layer on the input x, that is, the output value of the fully connected layer, and W base is the basis weight, σ(x) is the activation function of the basis weight; W spline is the spline weight, B(x) is the basis function of the spline interpolation; k and s are the current node index and the order of the spline interpolation, respectively, grid k Represents the value of the kth grid node used for spline interpolation in the fully connected layer, grid k+s+1 Represents the k+s+1th grid node, which is used to calculate the distance between the current node and the next node. k+s+2 Indicates the k+s+2th grid node, which is used for higher index values ​​required in interpolation calculations. k+s-2 Represents the k+s-2th grid node, which is used to calculate earlier node values ​​and to construct the polynomial basis during interpolation. k+1 Represents the k+1th grid node, which is used in the calculation of the interpolation basis function to determine the boundary of the interpolation range; S34, the output layer uses the Sigmoid activation function to generate the predicted probability y of students mastering the knowledge points: y=σ(KLinear3(KLinear2(KLinear1(x)))) Based on the students' knowledge mastery distribution, we can obtain a student portrait that reflects their learning status, including the mastery level and progress of each knowledge point.

5. According to claim 4, a personalized learning resource recommendation method based on subject knowledge graph and graph neural network is characterized in that: In step S4, the following steps are included: S41. Use graph embedding method to vectorize each knowledge point entity in the knowledge graph: vectorize the knowledge point k and the relationship between the knowledge point and other knowledge points r ij This is processed through a language model, which encodes the input text description into an embedded representation: e k =BERT(T k ) Where, T k is the text description of knowledge point k; It is a relationship ij Text description, BERT () is a pre-trained bidirectional encoder representation converter model used to convert input text into an embedding vector; S42. Integrate the knowledge graph embedding result into the student portrait, including: based on the relational structure in the knowledge graph, use the graph neural network to propagate the semantic information of the knowledge point, and integrate the upstream and downstream information of the relevant knowledge point into the current student portrait; perform feature splicing on the student's mastery of the knowledge point and the vector embedded in the knowledge graph to form a more complete representation of the student's knowledge state, wherein, combined with the neighborhood aggregation operation of the graph neural network, the embedding representation of the knowledge point and the relationship is integrated into the student portrait, and the student portrait embedding after knowledge enhancement is obtained: For each knowledge point node k, collect its neighboring nodes The embedded information of , the neighborhood information is integrated through the aggregation function f: Where W (l) is the weight matrix of the lth layer, σ( ) is the activation function, concat( ) is the concatenation function, and f( ) is the aggregation function, which aggregates the embedding information of node k and its neighboring nodes. is the embedding information of the t-th domain node of node k in the l-th layer, and Respectively represent the embedding information of node k in layer l and layer l+1; The knowledge graph information after graph embedding is used to supplement the background knowledge of the student portrait and identify the mastery relationship of students on knowledge points.

6. According to claim 5, a personalized learning resource recommendation method based on subject knowledge graph and graph neural network is characterized in that: In step S5, the personalized recommendation model adopts a hybrid recommendation method to recommend learning resources by combining collaborative filtering recommendation based on group characteristics and content recommendation based on semantic characteristics. ik The calculation method is: In the formula, is the recommendation degree of knowledge point k given by collaborative filtering to student i is the recommendation degree of knowledge point k given by the content recommendation method to student i, and α is the weight parameter used to balance the influence of the two recommendation methods; Based on the output of the personalized recommendation model, personalized learning content is pushed to students. The learning content includes: knowledge point materials, specific exercises and learning paths.

7. The personalized learning resource recommendation method based on subject knowledge graph and graph neural network according to claim 6 is characterized by: In the personalized recommendation model, the recommendation process of collaborative filtering recommendation based on group characteristics is as follows: Based on the students’ knowledge and interest characteristics, we recommend other knowledge points that are similar to their known knowledge points. We calculate the similarity by vectorizing the knowledge points and students’ learning records: In the formula, i, j represent different students respectively; R ik , R jk Respectively represent the scores of student i and student j on knowledge point k, and n is the total number of knowledge points; Predict the recommendation degree of student i for unlearned knowledge point k based on similarity In the formula, r i is the average mastery of student i, r jk is the mastery of knowledge point k by student i, and N(i) is the set of students most similar to student i.

8. According to claim 6, a personalized learning resource recommendation method based on subject knowledge graph and graph neural network is characterized by: In the personalized recommendation model, the recommendation process of content recommendation based on semantic features is as follows: Use vector representation and semantic matching to achieve semantic enhancement of knowledge graph. Suppose knowledge point k is embedded in the graph embedding model of knowledge graph to obtain embedding vector v k , and the embedding vector of the student portrait is v i , the semantic similarity is calculated as follows: Where D is the vector dimension, v id Represents the student portrait embedding vector v i The d-th dimension vector in, v kd Represents the knowledge graph embedding vector v k The d-th dimension vector in ; Based on the semantic similarity, the recommendation degree of student i for knowledge point k is predicted In the formula, β is a regulation parameter used to adjust the impact of content recommendation on the overall recommendation degree.

Images